Compiler Construction Lecture 16 Data-Flow Analysis

Goal of Data-Flow Analysis Automatically compute information about the program Use it to report errors to user (like type errors) Use it to optimize the program Works on control-flow graphs: x = 1 while (x < 10) { x = x + 2 }

How We Define It Abstract Domain D (Data-Flow Facts): which information to compute? – Example: interval for each variable x:[a,b], y:[a’,b’] Transfer Functions [[st]] for each statement st, how this statement affects the facts – Example:

Find Transfer Function: Plus If and we execute x= x+y then Suppose we have only two integer variables: x,y So we can let a’= a+c b’ = b+d c’=c d’ = d

Find Transfer Function: Minus If and we execute y= x-y then Suppose we have only two integer variables: x,y So we can let a’= a b’ = b c’= a - d d’ = b - c

Transfer Functions for Tests if (x > 1) { y = 1 / x } else { y = 42 }

Merging Data-Flow Facts if (x > 0) { y = x } else { y = -x – 50 } join

Handling Loops: Iterate Until Stabilizes x = 1 while (x < 10) { x = x + 2 } Compiler learned some facts!

Data-Flow Analysis Algorithm var facts : Map[Vertex,Domain] = Map.withDefault(empty) facts(entry) = initialValues // change while (there was change) pick edge (v1,statmt,v2) from CFG such that facts(v1) was changed facts(v2)=facts(v2) join [[statmt]](facts(v1)) } Order does not matter for the end result, as long as we do not permanently neglect any edge whose source was changed.

Handling Loops: Iterate Until Stabilizes x = 1 while (x < 10) { x = x + 2 } Compiler learned some facts!

Handling Loops: Iterate Until Stabilizes x = 1 while (x < n) { x = x + 2 } Compiler learns some facts, but only after long time n =

Handling Loops: Iterate Until Stabilizes var x : BigInt = 1 while (x < n) { x = x + 2 } For unknown program inputs it may be practically impossible to know how long it takes var n : BigInt = readInput() Solutions - smaller domain, e.g. only certain intervals [a,b] where a,b in {-∞,-127,-1,0,1,127,∞} - widening techniques (make it less precise on demand)

Size of analysis domain Interval analysis: D 1 = { [a,b] | a ≤ b, a,b  {-M,-127,-1,0,1,127,M-1}} U {  } Constant propagation: D 2 = { [a,a] | a  {-M,-(M-1),…,-2,-1,0,1,2,3,…,M-1}} U {  } suppose M is 2 63 |D 1 | = |D 2 | =

How many steps does the analysis take to finish (converge)? Interval analysis: D 1 = { [a,b] | a ≤ b, a,b  {-M,-127,-1,0,1,127,M-1}} U {  } Constant propagation: D 2 = { [a,a] | a  {-M,-(M-1),…,-2,-1,0,1,2,3,…,M-1}} U {  } suppose M is 2 63 With D 1 takes at most steps. With D 2 takes at most steps.

Termination Given by Length of Chains Interval analysis: D 1 = { [a,b] | a ≤ b, a,b  {-M,-127,-1,0,1,127,M-1}} U {  } Constant propagation: D 2 = { [a,a] | a  {-M,…,-2,-1,0,1,2,3,…,M-1}} U {  } U {T} suppose M is 2 63 Domain is a lattice. Maximal chain length = lattice height

Lattice for intervals [a,b] where a,b  {-M,-127,0,127,M-1}

Lattice

Data-Flow Analysis Ingredients

Transfer Function Pre-state Post-state S

Abstraction/Concretization Abstraction Function   Concretization Function 